Do Zero-Shot Dynamics Models Assume the World Is Deterministic?

Pretrained forecasting models for dynamical systems now extrapolate to equations they have never seen. Chronos, Panda, DynaMix and the Foundation Inference Model all match or beat models trained on the target system. Every one of them was trained, evaluated, or both on clean deterministic equations: the next state is a fixed function of the current one. Real systems also have randomness, and randomness comes in two kinds that look the same in a single noisy trace but behave very differently going forward. Measurement noise is error added to the readout; it never feeds back into the dynamics. Process noise is random forcing that pushes the system itself and compounds over time. A model that has quietly absorbed the deterministic assumption should treat both as nuisance to smooth away. It will therefore misjudge how fast uncertainty grows, and it will miss behaviour that only randomness creates, such as a particle hopping between two wells that the noiseless equations say it never leaves.

We propose a zero-shot benchmark built from matched pairs of stochastic systems with known ground-truth futures. In each pair, the same deterministic core is corrupted by either measurement noise or process noise, tuned so the observed traces have identical variance. The benchmark scores whether a pretrained model's predicted distribution of futures matches the true one. We also run one controlled intervention: the same architecture pretrained on deterministic versus stochastic equations, with the number of distinct systems held fixed. Together these tell us whether the failure, if present, is a property of the pretraining data or of the forecasting mechanism itself. The test that matters for downstream use is whether the models can estimate, zero-shot, the probability that a system leaves a safe region within a horizon. That quantity is what probabilistic safety tools in control actually need.

Core Claim

Current zero-shot dynamics models encode an implicit determinism prior. Concretely, given contexts with equal observed variance, their forecast spread will grow at nearly the same rate whether the noise is measurement or process noise, whereas the true spread differs. They will also systematically underpredict noise-induced transitions and therefore underestimate escape probabilities. We further claim the gap is mostly a data effect. Pretraining on stochastic differential equations (SDEs), which are equations of motion with an explicit random-forcing term, should close it while keeping zero-shot transfer. The alternative outcome is that the gap is structural, because the dominant mechanism is in-context copying of past segments, and copying cannot invent a transition it has not seen. Either outcome is informative, and the experiments are designed to tell them apart.

Relation to Prior Work

Four papers motivate the proposal by showing strong zero-shot transfer on deterministic systems while leaving stochasticity untested.

  • Zero-shot forecasting of chaotic systems probes Chronos on 135 chaotic systems from the dysts collection. Chronos reaches roughly one Lyapunov time of valid prediction (the timescale over which small errors in a chaotic system grow by a factor of e) and preserves attractor shape. The authors attribute this to "context parroting", the copying of similar past segments, and note that it breaks when a real pendulum loses energy.
  • Panda pretrains on 20,000 synthetic chaotic equations. It finds that performance scales with the number of distinct systems rather than raw data volume, a finding we reuse to design a fair pretraining ablation.
  • DynaMix shows that a roughly 10,000-parameter mixture of recurrent experts preserves long-run statistics zero-shot, but it assumes stationary, autonomous, deterministic flows.
  • The Foundation Inference Model imputes gaps by inferring an underlying ordinary differential equation, and so builds the deterministic assumption directly into its architecture.

None of these four distinguishes the two noise sources or scores predicted distributions against a true stochastic future. That is the specific thing we add.

Two lines from control constrain and motivate the safety check.

  • ADAPT handles stochastic disturbances at transfer time with a hand-built tube controller around one policy. It does not ask what a general pretrained model knows about noise.
  • Work on stochastic control barrier functions and data-driven certificate synthesis shows that safety under noise is a statement about probabilities computed from samples. That makes sample quality, the thing we measure, directly relevant.

We do not synthesize certificates. We test whether foundation-model samples are trustworthy enough to feed such tools.

Research Program

Benchmark construction. We take low-dimensional drift equations from dysts and from ODEBench and build three noise regimes for each, using the Euler–Maruyama method (the standard step-by-step simulation scheme for SDEs):

  • additive measurement noise only;
  • additive process noise only;
  • state-dependent process noise.

Noise amplitudes are calibrated per system so that the stationary variance of the observed signal matches across regimes to within 5%. This removes "the trace looks noisier" as an explanation for any difference in model behaviour.

Noise-induced phenomena. We add a second family of systems whose qualitative behaviour exists only because of noise: double-well potentials with escape events, excitable systems that fire under noise but rest without it, and stochastic resonance. In these, a deterministic-prior forecast predicts something qualitatively wrong rather than just imprecise.

Ground truth. For every context window we draw 10,000 Monte Carlo continuations from the true SDE given the true latent state. This yields the exact conditional future distribution and exact first-exit probabilities from a designated safe set.

Metrics. We use four.

  1. Continuous ranked probability score and energy score. Both are proper scoring rules, meaning they reward a predicted distribution for being close to the true one and cannot be gamed by hedging.
  2. Calibration of predicted quantiles.
  3. A spread-growth discrepancy: the ratio of a model's forecast spread under process noise to its spread under measurement noise, compared with the true ratio. A pure determinism prior produces a ratio near one.
  4. Absolute error in the estimated probability of leaving the safe set within horizon $H$.

Intervention. We pretrain one encoder architecture in Panda's style twice, on 5,000 systems each, matched in system identity and token count. The first corpus is deterministic. The second gives each system randomly sampled process and measurement noise. Both models are evaluated zero-shot on held-out systems.

Baselines

Naive reference points. The first is climatology: the empirical distribution of the context window, repeated forward. It has no dynamics, so any model that cannot beat it on short horizons has learned nothing about how this particular system evolves. The second is persistence plus noise: the last observation with Gaussian spread fitted from context.

Analog ensemble. This is the most dangerous simple competitor. It makes context parroting explicit: find the $k$ nearest past segments in a delay embedding of the context and use their continuations as forecast samples. It is the mechanism Zero-shot forecasting of chaotic systems identified behind Chronos's success, and it naturally produces a distribution. It is tuned by choosing $k$ and the embedding dimension on a disjoint validation set of systems, then fixed.

Existing pretrained models. These are Chronos (sampled forecasts), Panda and DynaMix, with spread for the latter two obtained by perturbing the context with noise estimated from the context itself. All run zero-shot with identical context lengths and horizons.

The strongest reported result in the file is Chronos's valid prediction time of approximately one Lyapunov time on deterministic dysts. No reported result was found on Lacuna for any of these models on stochastic dynamics or on distributional scores. Beating the field therefore means two things:

  • The stochastically pretrained model lowers continuous ranked probability score relative to the best of these baselines on held-out systems in both noise regimes.
  • It brings the spread-growth ratio within 20% of the true ratio where the baselines are near one.

Oracles. The upper oracle is the true SDE simulated from a filtered state estimate. A per-system supervised model trained on long target trajectories gives the "just train on it" reference.

Ablation. The deterministic-versus-stochastic pretraining pair isolates the key ingredient: exposure to noisy dynamics, with system diversity held fixed.

Why not just…?

Why not just add measurement noise to dysts and rerun Chronos? That is half of our benchmark, but on its own it cannot detect a determinism prior. A model that smooths everything would do well on measurement noise. The diagnostic signal comes only from the matched comparison with process noise at equal observed variance, which this alternative omits.

Why not just use an analog ensemble, since parroting is what the models do anyway? This is a real threat on stationary systems with long contexts, where analogs sample the attractor honestly. Copying, however, can only reproduce transitions present in the context. On double-well systems where the context contains no escape, it should assign near-zero escape probability. The noise-induced family is built to show exactly this. If the analog ensemble matches the pretrained models there too, we will report that the models add nothing beyond copying.

Why not just augment pretraining with noise and skip the benchmark? That augmentation is our intervention. Without matched-pair evaluation and Monte Carlo ground truth, though, there is no way to know whether it helped calibration or merely widened forecasts. Widening wins on some scores while still being wrong about how spread grows over time.

Decisive Checks

The first check is whether existing models show a spread-growth ratio near one across regimes on held-out dysts drifts. If they do not, the determinism-prior hypothesis is false and we say so.

The second is the escape-probability experiment on noise-induced systems. Here the stochastically pretrained model must reduce absolute exit-probability error relative to both the deterministic twin and the analog ensemble.

The third is a real-data check on the video-tracked double pendulum from Zero-shot forecasting of chaotic systems. We bootstrap held-out segments, and the question is whether better-calibrated synthetic performance carries over to real damping and real noise.